Same. The female voice, especially on the first sentence in the cast, is very well inflected to separate phrases and add interest. After that, it's downhill.
ps, I feel like inflection is going to be one of the harder things for an LM to pick up, given all the subtext humans can convey with it.
Laura sounds very realistic... Zod is a bit less so . Both still very impressive. This was really cool. I'm excited to see all the new ideas with this api access.
Realistic but very lacking in expression and no humor at all. And very slow paced. I'd want significantly more personality to be happy with it - I wonder if the reason its like this is because when they try to spice it up we are back to inappropriate things popping out.
Looks like automated news is finally achieved. I remember in the early 2000's how I became impressed by Ananova and it wasn't even close to fully automated. This one seems to work really well.
According to the podcast itself it runs on Azure, so very likely it's Azure TTS. I also think that's somewhat evident because Elevenlabs TTS is (at least in my opinion) a bit more natural than Azure TSS.
I am still searching for a good open-source library that produces natural voices. I have experimented with Coqui-ai and Mimic 3, but they are not this good. I have heard that Tortoises-tts is quite slow.
I would love to know about any other alternatives that I may have missed.
Imagine this in a future GTA game where the news loop is closed and self generating. Endless radio content and commentary based on havoc in the city, winning online gambles etc.
```
That's all for the weather report. Now we have some breaking news. A maniac has stolen a tank from a military base and is leading the police on a wild chase through the city streets. We have our reporter on the scene with more details. Stay tuned for this developing story.
```
Pre-programmed common stuff sure which many games do as they can tell where you are, vehicle in and any weapons used etc, steal a (police car|military jet|ambulance) etc and they can craft scenarios using audio for those but for more natural random somewhat unpredictable stuff you would have endless combinations that need to be accounted for. It would have to act in response to a bunch of actions so not feasible.
A good example is Mortal Kombat where each character has several lines for every OTHER character (https://www.youtube.com/watch?v=L85QApISlvA) they are about to fight, that's a LOT of voice work as opposed to using the character lore and their history and or relationship with the opponent to come up with something fresh that's (witty|snarky|sad) to say etc.
Yes but this stuff is taking 100 gigabytes per game
Once we get flexgen running Llama (or some other combination of optimizations) there will be no audio files and these LLM’s can be run client side on consumer hardware
Exactly, imagine every NPC just has some ai generated backstory, personality traits, dislikes, habits etc and then these interact with some other NPC or group with same and they use these to have natural flowing convos. Now imagine debates with groups that learn and have the possibility to "evolve", I'm getting a mix of Red Dead Redemption and Dwarf Fortress here for content, the possibilities are endless.
This is pretty wild. Eerie how relatable the hosts are, talking about where they’re from, etc. There is an uncanny valley feel to it though. For example, Laura said GitHoob breaking the “illusion”.
The one that’s driven me crazy lately is when Siri tells me through my AirPods that I left something beind. It always pronounces the “St.” in the address as “saint” instead of “street”, and I can’t understand how it would do this by accident.
There is a beautiful bit in Little, Big by John Crowley. At the start of the book one of the characters is working checking entries in the telephone directory and is amused that the system has confused saint and street to produce Church Of All Streets and the Seventh Saint Bar. Later in the book both locations are mentioned, and it turns out were correctly named in the telephone directory.
Is there a reason the voices are so slow? This is even slower than people who are trying to talk slow, and it feels so out of place... there is the speed setting, and 1.2x makes the speech sound way more like an actual human.
Kind of like right now with 90% of all mainstream media being owned by just 6 corporations. Their employees must abide by the rules they set and are told what they can and cannot talk about.
I'd venture to say, this will only increase people's skepticism, which is a good thing. We need people to start thinking for themselves instead of turning off their brain and just being fed info they assume they can trust.
I hope that's true, but I suspect the allure of getting your own personalized news feed on only the topics you care about, in the exact style you prefer will cause 90% of the listeners to choose this medium over all others and presume as much (or more) truth in this than the source they prefer today.
Never discount the influence of high production value on any form of media. Look at the utter crap music and films that have dominated mass media for decades. The best produced and most palatable fare nearly always sells best, no matter what the quality of the underlying content.
End of the road for podcasts more like. They are incredibly labour intensive to produce (recording + editing time), and more and more of them are becoming not much more than plugs for their book, TV show or what not. I can see them turning the medium into an automated marketing channel, the way email lists are today.
This is technically very impressive, but it's worth pointing out that podcasts much better than this fail to build an audience all the time.
I also feel like every application of ChatGPT seems to completely miss the point of the media it mimics. Podcasts are not merely coherent voices talking to each other. Getting rid of human presenters is literally soulless. People already don't listen for much subtler reasons. Entertainers get canceled, media companies get boycotted, bias divides audiences, etc.
That's not going away with or without AI. There is no "tweaking" the training without putting humans right back into the equation and probably making production way more expensive than it's worth. There is no scalability payoff either. Who wants to listen to the same podcast cloned a million times with just replaced voices? We already have this problem with podcasts today and it kills any interest to consume it.
The scalability payoff is in personalization. E.g. I love "This week in microbiology", but I wish I could have more influence over the scientific papers discussed. What I'd love is a morning podcast that's exactly as long as I eat breakfast that talks about exactly the papers I'm reading and their interconnections.
No, an echo chamber is a space without dissenting opinion.
Personalisation could be used to make an echo chamber, but to branch off the microbiology example above, personalisation of content could also be a summary of all the debate happening in the niche.
I think it would be quite a bit more interesting if you could converse with the model. The back and forth "is this paper about foo related to this other paper about bar?" would probably be a better way of getting at the interconnections. This should be doable now.
The thing that might hold it back is the latency in the experience. You could mask it with the AI equivalent of "ummm ..." to get to maybe 5-10s.
Yes, but would you really love a morning podcast that's
* exactly as long as your breakfast consumption time
* talks about the papers you're reading, but...
* is as shallow as a puddle and as funny as being the person who steps in one?
Because that's what this is. The synthesized discussion combines all the insight of a breakfast radio host interviewing a guest on a specialized technical topic, and the banter as engaging as a technical specialist of some kind trying to host breakfast radio.
By the way, I'm not trying to be overly critical of the developers of this experiment, which is a great illustration of where we're currently at with a bunch of technologies. But it also very starkly illustrates its current limitations.
The purpose of a podcast (for me) isn't just to curate content (as this is doing), but to get the perspective of the individual domain experts hosting the show. AI can't address that key motive until it produces models whose particular opinions and analysis I want to hear about topics I probably have already found elsewhere.
Right now, people are in the "This is really cool" phase of using the technology. People are learning to use it by implementing whatever strikes their fancy, including a lot of things that weren't possible before, but which aren't practical or valuable.
Once things settle down we'll start to see some seriously useful stuff, but for the moment it's the wild west.
I def agree with what you're saying, and so this is definitely not for me, but part of me wonders if this might become the next generational divide (ie if kids grow up with this type of content normalized, maybe they don't react as negatively?).
The scenario I fear most would be AI generated opinionated podcasts aimed at humans, with the purpose of directing their preference, that is, "propagandAI". This already happens daily with traditional media, but that also gives us the weapons to fight it because there's a person on the other side and we know humans can be evil or just fail. But who is to blame when million of people put in power the wrong person because of what an AI that is not a legal person, still they deeply trust because "machines can't lie", directed them to by pushing the right buttons in their heads?
What concerns me the most isn't the AI itself but rather the humans behind it that will use it to take advantage of other humans.
I'm not sure about "podcasts" but this concept could be for sure used in news channels, as we have for example in Germany, hourly. It would for sure save money from our taxpayers.
Laura and Zod sound remarkably similar to the narrators in this audible I recently listened to called After On: A Novel of Silicon Valley (not recommending it!) and I seriously wonder if the whole book wasn't narrated by AI.. it's not the first audible that made me wonder.
It's funny how these two can talk about "starship bridge simulators" or "gnu poke" like they are super enthusiasts. I think one of the key personality characteristics of ChatGPT is its endless enthusiasm for stuff that can be incredibly geeky, niche, weird or boring to most people.
"Sounds like super useful pickles for those who work with binary files!"
161 comments
[ 0.32 ms ] story [ 202 ms ] threadIf I wasn't clued in, I probably wouldn't know these weren't human. At least the male voice sounds slightly more natural to me.
ps, I feel like inflection is going to be one of the harder things for an LM to pick up, given all the subtext humans can convey with it.
I thought so too. What do they use for text to voice?
> The hosts are Laura and Zod
https://hackerfm.com/about
The voice sounds uncanny with unusual breathing pauses, and there isn’t a name announced when they come on or sign off the traffic report.
(1) https://jazz.fm/
I would love to know about any other alternatives that I may have missed.
I want Sam Jackson and Molly Wood to read my hn please.
I would really like to have a timestamp to click in the story listing.
This would begin playing the audio at that story.
Well... OpenAI did also release the Whisper API
``` That's all for the weather report. Now we have some breaking news. A maniac has stolen a tank from a military base and is leading the police on a wild chase through the city streets. We have our reporter on the scene with more details. Stay tuned for this developing story. ```
A good example is Mortal Kombat where each character has several lines for every OTHER character (https://www.youtube.com/watch?v=L85QApISlvA) they are about to fight, that's a LOT of voice work as opposed to using the character lore and their history and or relationship with the opponent to come up with something fresh that's (witty|snarky|sad) to say etc.
Once we get flexgen running Llama (or some other combination of optimizations) there will be no audio files and these LLM’s can be run client side on consumer hardware
probably as a shared resource at the OS level
Someday the AI will introduce mistakes on purpose to seem more human like.
Will be very difficult to detect in the future and will result in trust issues / rampant fake news.
I'd venture to say, this will only increase people's skepticism, which is a good thing. We need people to start thinking for themselves instead of turning off their brain and just being fed info they assume they can trust.
https://www.businessinsider.com/these-6-corporations-control...
Never discount the influence of high production value on any form of media. Look at the utter crap music and films that have dominated mass media for decades. The best produced and most palatable fare nearly always sells best, no matter what the quality of the underlying content.
If I could choose a preference for personality and voice, I'd probably be sold.
Any affiliation with https://old.reddit.com/r/airadio/ ?
I also feel like every application of ChatGPT seems to completely miss the point of the media it mimics. Podcasts are not merely coherent voices talking to each other. Getting rid of human presenters is literally soulless. People already don't listen for much subtler reasons. Entertainers get canceled, media companies get boycotted, bias divides audiences, etc.
That's not going away with or without AI. There is no "tweaking" the training without putting humans right back into the equation and probably making production way more expensive than it's worth. There is no scalability payoff either. Who wants to listen to the same podcast cloned a million times with just replaced voices? We already have this problem with podcasts today and it kills any interest to consume it.
EDIT: to be clear I'm not hating on LLMs, but that whatever the next big thing is probably won't be imitating what exists today
Personalisation could be used to make an echo chamber, but to branch off the microbiology example above, personalisation of content could also be a summary of all the debate happening in the niche.
The thing that might hold it back is the latency in the experience. You could mask it with the AI equivalent of "ummm ..." to get to maybe 5-10s.
* exactly as long as your breakfast consumption time
* talks about the papers you're reading, but...
* is as shallow as a puddle and as funny as being the person who steps in one?
Because that's what this is. The synthesized discussion combines all the insight of a breakfast radio host interviewing a guest on a specialized technical topic, and the banter as engaging as a technical specialist of some kind trying to host breakfast radio.
By the way, I'm not trying to be overly critical of the developers of this experiment, which is a great illustration of where we're currently at with a bunch of technologies. But it also very starkly illustrates its current limitations.
Once things settle down we'll start to see some seriously useful stuff, but for the moment it's the wild west.
A possible use case for this could be podcasts dealing with inflammatory, politically divisive topics and disguised as coming from real hosts.
"Sounds like super useful pickles for those who work with binary files!"